howdy, it’s Barsee again.

happy friday, AI family, and welcome back to AI Valley.

here are the biggest things worth knowing today:

• OpenAI launches new realtime voice and translation AI models

• AI researchers are betting on “world models” instead of chatbots

• Plus trending AI tools, posts, and resources

Let’s dive into the Valley of AI…

FLOW

Speak naturally. Send without fixing.

Wispr Flow turns your voice into ready-to-send text in any app. Give coding agents 10x more context by speaking instead of typing. Works in Cursor, Claude, ChatGPT, and every IDE.

89% of messages sent with zero edits. Millions of users worldwide. Available on Mac, Windows, iPhone, and now Android.

Free and unlimited on Android during launch.

Try Flow Free

_{*This is sponsored}

THROUGH THE VALLEY

1/ OpenAI launches new realtime voice and translation AI models

OpenAI introduced three new realtime audio models through its API platform: GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper. (see the demo here)

The release is aimed at developers building voice agents, live translation tools, and streaming transcription products rather than regular ChatGPT users.

GPT-Realtime-2 is the main upgrade. OpenAI says the model can handle more complex requests, recover from mistakes during conversations, use multiple tools at once, and maintain more natural back-and-forth interactions in real time.

The model’s context window also increased from 32K to 128K tokens, allowing it to retain much longer conversations and customer histories during calls.

OpenAI also launched GPT-Realtime-Translate for multilingual conversations across 70+ languages and GPT-Realtime-Whisper for realtime speech transcription.

Companies including Zillow, Priceline, Vimeo, and Deutsche Telekom are already testing the models for customer support, travel assistance, live translation, and voice-based workflows.

Why this matters:

Voice AI has struggled with a tradeoff for years: systems could either respond quickly or reason well, but rarely both. Faster models often sounded natural but made mistakes, while smarter models introduced long pauses that broke conversations. OpenAI is trying to close that gap by making voice systems think and respond in real time without awkward delays. The bigger opportunity is that voice could eventually become a major interface for software itself, especially in areas like support, scheduling, healthcare, sales, and travel where people still spend hours talking instead of typing.

2/ AI researchers are betting on “world models” instead of chatbots

A growing number of AI researchers believe language models alone won’t be enough to reach real-world intelligence.

Companies like Yann LeCun’s AMI Labs, Fei-Fei Li’s World Labs, and Skild AI are now building “world models,” systems trained to understand how the physical world works through video, simulations, and real-world interactions.

The idea is that chatbots learned language from internet text, while world models would learn things like movement, physics, cause and effect, and how objects behave in real environments.

Some researchers argue current AI still lacks basic physical understanding because it only learns from text instead of direct experience.

That limitation keeps appearing in newer ARC-AGI tests designed to measure reasoning in unfamiliar situations. OpenAI’s o3 scored 87.5% on the original benchmark, but newer versions reportedly pushed top models below 1% while remaining easy for humans.

Why this matters:

Current AI is very good at predicting language but much weaker at understanding how the real world works. That becomes a major problem once systems move into robotics, automation, and real-world decision-making.

The challenge is that physical-world data is far harder to collect than internet text. Robots need huge amounts of examples showing how objects move, break, collide, and react. If companies solve that data problem, world models could become one of the next major shifts in AI.

TRENDING TOOLS

• Google > Launched its AI health coach publicly after months in beta, combining Fitbit, Health Connect, wearable data, and U.S. medical records

• Reactor > Experience entire worlds generated in real time on a global low-latency infrastructure

• Higgsfield Ad Reference > Upload your best-performing videos and it automatically recreates the format

• OpenAI Voice Models > OpenAI launched three new voice models: GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper

• FlowMarket > Orchestrates a B2B network where autonomous agents use real-time algorithmic matching to automate discovery and engagement

• Kanwas > A collaborative context brain that turns team knowledge into a spatial reasoning and delivery workspace

• Claude for Microsoft 365 > Now generally available inside Excel, PowerPoint, Word, and Outlook (beta)

• Save to Spotify > Spotify’s new command-line tool lets AI agents upload AI-generated podcasts directly to the platform

WHAT I'M CONSUMING

• Is AI 2027 coming true?

• How AI agent memory works

• You are not a horse

• OpenAI adds emergency contacts to ChatGPT for mental health concerns

THE VALLEY GEMS

What’s trending on social today:

THAT’S ALL FOR TODAY

Thank you for reading today’s edition. That’s all for today’s issue.

💡 Help me get better and suggest new ideas at [email protected] or @heyBarsee

👍️ New reader? Subscribe here

Thanks for being here.

REACH 100K+ READERS

Acquire new customers and drive revenue by partnering with us

Sponsor AI Valley and reach over 100,000+ entrepreneurs, founders, software engineers, investors, etc.

If you’re interested in sponsoring us, email [email protected] with the subject “AI Valley Ads”.